MCP child processes orphaned on restart — memory leak

[Cipher208]

Description

When MiMoCode restarts (crash, manual restart, or session end), child MCP processes are NOT terminated. They get reparented to PID 1 (systemd/init) and persist as orphans, consuming memory indefinitely.

Each restart leaks 2 processes (~81MB RAM). After multiple restarts in a session:

• 22 orphaned shell+MCP pairs = 1.7GB RAM wasted
• 47 total MCP processes visible (only 2 legitimate)

Reproduction

1. Configure MCP servers in mimocode.json (e.g., ariel-memory with Python stdio)
2. Start MiMoCode
3. Restart MiMoCode 10+ times (e.g., during plugin/skill development)
4. Check: ps aux | grep mcp_server | wc -l → shows 20+ processes
5. Check orphans: ps -eo pid,ppid,comm | grep 'python3.*mcp_server' | awk '$2==1' → lists orphans

Root Cause Analysis

MiMoCode spawns MCP servers via shell wrapper:

sh -c 'cd /path && python3 -m mcp_server.server --transport stdio'

Process chain: MiMoCode (Node.js) → sh (shell wrapper) → python3 (MCP server)

When MiMoCode exits/restarts:

1. Node.js process dies
2. Shell wrapper (sh) reparented to PID 1 (systemd)
3. Python MCP server also reparented to PID 1
4. New MiMoCode instance creates new MCP processes
5. Old processes remain as orphans indefinitely

Each orphan pair costs ~81MB RAM (1MB shell + 80MB Python MCP server).

Expected Behavior

MCP child processes should be terminated when MiMoCode exits or restarts. Options:

• Use process group killing (setsid + kill -- -PGID)
• Track child PIDs and kill them in shutdown handler
• Use prctl(PR_SET_PDEATHSIG, SIGTERM) in the MCP launcher so children die when parent dies

Actual Behavior

MCP child processes become orphans (PPID=1) and consume ~80MB each until manually killed or system reboot.

Environment

• MiMoCode version: v0.38.9
• OS: Debian 12 (Linux)
• Node.js: system Node
• MCP config: ariel-memory (Python, stdio), context-mode (Node, stdio)
• Install method: npm global (npm install -g mimocode)

Workaround

Cron job to clean up orphaned MCP processes every 15 minutes:

#!/bin/bash
# Kill mcp_server processes orphaned to PID 1
ps -eo pid,ppid,comm | grep 'python3.*mcp_server' | awk '$2==1 {system("kill -9 " $1)}'
ps -eo pid,ppid,args | grep 'sh.*-c.*mcp_server' | awk '$2==1 {system("kill -9 " $1)}'

Impact

On a development VPS with frequent MiMoCode restarts (plugin development, debugging), this leaks ~1.7GB RAM per session. Without the workaround, the system runs out of memory within hours.